Session Reconstruct

Session Reconstruct

Retroactively analyze and annotate exported sessions to reveal orchestration that wasn't captured.

Note: This skill analyzes sessions exported via the built-in /export command or raw JSONL logs from ~/.claude/projects/. It INFERS orchestration details that weren't narrated—accuracy is ~60-80% vs ~95% for --showcase mode.

Quick Start

# For current session (export + reconstruct in one step)
"Export and reconstruct this session --reconstruct"

# For already-exported file
"Reconstruct orchestration from session.md --reconstruct"

# Other options
"Analyze this session --audit"
"Walk through what happened --replay"

Important: /export --reconstruct won't work because /export is a built-in command that doesn't accept flags. Use the natural language commands above instead.

For NEW sessions, use showcase-export with --showcase instead.

How It Works

┌─────────────────────────────────────────────────────────┐
│ Input Sources                                           │
├─────────────────────────────────────────────────────────┤
│ 1. /export output (.md or .txt)                         │
│ 2. Raw JSONL logs (~/.claude/projects/*.jsonl)          │
│ 3. Community tool exports (claude-code-log, etc.)       │
└─────────────────────────────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────┐
│ Reconstruction Engine                                   │
├─────────────────────────────────────────────────────────┤
│ • Identifies skill invocations from output patterns     │
│ • Infers agent reasoning from results                   │
│ • Reconstructs decision points from choices made        │
│ • Estimates compound learning from behavior changes     │
└─────────────────────────────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────┐
│ Output: Annotated Session                               │
├─────────────────────────────────────────────────────────┤
│ Original transcript + [RECONSTRUCTED] markers           │
│ with confidence scores for each inference               │
└─────────────────────────────────────────────────────────┘

When to Use This

What Gets Reconstructed

1. Skill Logic (from outputs)

[RECONSTRUCTED SKILL LOGIC]
Skill: idea-validator
Based on the output pattern, this skill likely instructed:
1. Problem clarity analysis (evidence: "clear problem" in output)
2. Market need validation (evidence: reference to "demand signals")
3. Competitive moat assessment (evidence: "defensibility" section)
Confidence: 85%

2. Subagent Internals (from results)

[RECONSTRUCTED AGENT PROCESS]
Agent: rigorous-thinking
Final result mentioned: "4/5 counterarguments addressed"
Inferred process:
- Generated ~5 counterarguments (evidence: "4/5" ratio)
- Tested each against evidence (evidence: "addressed" language)
- Tool calls: ~4-6 (typical for this agent type)
Confidence: 70%

3. Decision Points (from choices made)

[RECONSTRUCTED DECISION]
At this point, the session chose X over Y.
Likely tradeoffs considered:
- X advantage: [inferred from context]
- Y advantage: [what was given up]
- Why X won: [reasoning based on subsequent actions]
Confidence: 60%

4. Compound Learning (from patterns)

[RECONSTRUCTED COMPOUND UPDATE]
A pattern was likely extracted here:
- Pattern: "[inferred from repeated behavior]"
- Evidence in session: [what suggested this]
- Likely confidence update: [estimate]
Confidence: 50%

Reconstruction Protocol

Step 1: Identify Orchestration Points

Scan for:

• Skill invocations (Skill:, 🔧, skill names mentioned)
• Agent spawns (Task, 🤖, "spawning", "agent")
• Phase transitions (numbered sections, "Phase", "Step")
• Decision indicators ("chose", "decided", "instead of", "rather than")
• Compound signals (database mentions, "pattern", "learned", "updated")

Step 2: Mark Confidence Levels

Step 3: Generate Annotated Version

# Session Reconstruction: [Project Name]

## Reconstruction Metadata
- Original session: [filename]
- Reconstruction date: [date]
- Overall confidence: [average %]
- Gaps identified: [count]

---

[ORIGINAL CONTENT]
User: Build sessionizer

[RECONSTRUCTION]
This request triggered the following orchestration:
- Skills likely loaded: idea-validator, software-architecture
- Why: "Build" keyword + project name suggests full build pipeline
- Confidence: 75%

Reconstruction Markers

Complete Workflow

# If you FORGOT --showcase:

# 1. Export the session using built-in command
/export my-session.md

# 2. Reconstruct orchestration using this skill
"Reconstruct orchestration from my-session.md --audit"

# 3. Output: Annotated version with [RECONSTRUCTED] markers

Comparison with showcase-export

Best practice: Always start with --showcase. Use --reconstruct only for old sessions or gaps.

Limitations

Reconstruction CANNOT provide:

1. Exact subagent reasoning - Can only infer from results
2. Precise tool call counts - Estimates only
3. Actual confidence scores - Must approximate
4. Internal decision debates - Only see final choice
5. Timing information - Unless explicitly logged

Always flag these limitations in the reconstructed output.

Installation

npx skills add sunnypatneedi/skills